International Journal of Computational Linguistics Research Djellali Hayet, Laskri Mohamed Tayeb 2 3 2011 http://www.dline.info/jcl/fulltext/v2n3_4/4.pdf We propose a new modeling approach in Automatic Speaker Verification A.S.V based on Gaussians Mixtures Models and Maximum a posteriori adaptation MAP. We propose clustering algorithm for intra and inter speaker’s variability in voice module and contribute for Universal Speaker Model design. We compare the traditional approach which uses one specific customer model with the second called Universal speaker model USM (customers families). Voice module is applied for characterizing customers only; Universal Speaker Model is applied when speaker model is weak and designed for computing a reliable score.